Contributions to High-Throughput Computing Based on the Peer-to-Peer Paradigm
Pérez Miguel, Carlos
MetadataShow full item record
This dissertation focuses on High Throughput Computing (HTC) systems and how to build a working HTC system using Peer-to-Peer (P2P) technologies. The traditional HTC systems, designed to process the largest possible number of tasks per unit of time, revolve around a central node that implements a queue used to store and manage submitted tasks. This central node limits the scalability and fault tolerance of the HTC system. A usual solution involves the utilization of replicas of the master node that can replace it. This solution is, however, limited by the number of replicas used. In this thesis, we propose an alternative solution that follows the P2P philosophy: a completely distributed system in which all worker nodes participate in the scheduling tasks, and with a physically distributed task queue implemented on top of a P2P storage system. The fault tolerance and scalability of this proposal is, therefore, limited only by the number of nodes in the system. The proper operation and scalability of our proposal have been validated through experimentation with a real system. The data availability provided by Cassandra, the P2P data management framework used in our proposal, is analysed by means of several stochastic models. These models can be used to make predictions about the availability of any Cassandra deployment, as well as to select the best possible con guration of any Cassandra system. In order to validate the proposed models, an experimentation with real Cassandra clusters is made, showing that our models are good descriptors of Cassandra's availability. Finally, we propose a set of scheduling policies that try to solve a common problem of HTC systems: re-execution of tasks due to a failure in the node where the task was running, without additional resource misspending. In order to reduce the number of re-executions, our proposals try to nd good ts between the reliability of nodes and the estimated length of each task. An extensive simulation-based experimentation shows that our policies are capable of reducing the number of re-executions, improving system performance and utilization of nodes.